Adversarial Examples Generation And Prediction Research Based On DQN Path Finding

As a new research hotspot in the field of artificial intelligence,deep reinforcement learning(DRL)has achieved certain success in various fields.At the same time,the possibility of its application being attacked or whether it have a strong resistance to strike has also become a hot topic in recent years.Therefore,we select the representative Deep Q Network(DQN)algorithm in deep reinforcement learning,and use the robotic automatic pathfinding application as a countermeasure application scenario for the first time,and attack DQN algorithm against the vulnerability of the adversarial samples.In this paper,we first use DQN to find the optimal path,the path finding path is the optimal and shortest path.At the same time,we analyze the rules of DQN pathfinding and evaluate the features of the path finding.We propose two methods to achieve the generation and prediction of adversarial samples.In the research of adversarial samples generation,by analyzing the two factors Q value and gradient values which affect the path planning of the DQN algorithm,we propose an effective method for generating the adversarial samples towards White-Box in DQN pathfinding training(WAG).Through this algorithm,it is possible to detect all adversarial sample points that may attack the path plan.These adversarial samples will interfere the robot pathfinding with different degree,making it impossible to achieve the optimal shortest path through the independent path finding and reduce its training efficiency.In the research of adversarial samples prediction,we propose an adversarial samples prediction method(APM).We analyze the features of all suspected adversarial samples found by the WAG algorithm,and according to the influence of the resistant sample points on the path,the resistant sample points are divided into two categories.Then,extract the features of their Q value and gradient value.The features are correlated and fused using the CCA algorithm.At the same time,two types of adversarial samples are tagged,we name the point that has the greatest impact on path planning as "fatal attack point",and the points other than this point are named "attack point".The KNN algorithm is used to predict the two categories of adversarial samples.In the experiment,the criteria for determining whether it is a confrontational sample is first formulated.Then a simulation environment is built as an experimental platform to test the method.Through a large number of experiments,we can successfully find the adversarial samples by WAG algorithm,and the accuracy of the classification prediction model established by the APM method reaches 94.8%.That means,the model can better predict the fatal attack sample points.